[Issue]: `error: operand #0 does not dominate this use`

xinyazhang commented 2 months ago

Problem Description

Regression found in a09c6b3e65ddde167266fc37e4f0b27c45ce0562 Cannot reproduce in 45d873225fe7f25e182edb7fd66fbae589114ad3 (a09c6b3e65ddde167266fc37e4f0b27c45ce0562~1)

Operating System

Ubuntu 20.04.6 LTS

CPU

AMD EPYC 7542

GPU

AMD Instinct MI210

ROCm Version

ROCm 6.0.0

ROCm Component

No response

Steps to Reproduce

git clone https://github.com/ROCm/aotriton.git
cd aotriton
pytest -s tritonsrc/test_backward.py -v -k 'test_op_bwd[False-1.2-dtype2-0.0-False-587-4-16-4-4]'

Docker Image: rocm/pytorch-private:compute-rocm-rel-6.1-82_ubuntu20.04_py3.9_pytorch_rocm6.1_internal_testing_70d92c0

(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

No response

Additional Information

No response

xinyazhang commented 2 months ago

Hit the following error when running the UTs shown above.

loc("/home/xinyazha/aotriton/tritonsrc/fwd_kernel.py":262:24): error: operand #0 does not dominate this use                                                                                                                                   
python: /home/xinyazha/triton/lib/Dialect/TritonGPU/Transforms/AMDReorderInstructions.cpp:130: void TritonAMDGPUReorderInstructionsPass::moveImmediatelyAfterOperands(mlir::Operation*, llvm::SmallVector<mlir::Operation*>&): Assertion `succ
eeded(mlir::verify(m))' failed.                                                                                        
Fatal Python error: Aborted

jerryyin commented 1 month ago

@oplavsic Have we already fixed this?

@xinyazhang Does this still exist in the latest triton upstream?

ROCm / triton